distinguish intermediate layer & prediction layer distillation phases with different parameters

[MichelBartels]

Currently, you have to use the same learning rate and training data for both stages of tinybert distillation. This behaviour does not match the original TinyBERT paper.

Proposed changes:
Add tinybert_learning_rate and tinybert_train_filename parameters to distil_from method of FARMReader.

Status (please check what you already did):

• First draft (up for discussions & feedback)
• Final code
• Added tests
• Updated documentation

[julian-risch]

I wonder if we can distinguish the two distillation techniques we now have a bit better. Rather than calling one technique tinybert and passing additional parameters to the same method distil_from method, we should think about other options. That includes a draft of how to describe the two options in a documentation guide. Maybe we can have two separate def distil_from() methods where one calls the other or one for 1st stage and one for 2nd stage distillation. Let's discuss in a call.

[julian-risch]

With the two separate methods distil_intermediate_layers_from and distil_prediction_layer_from it's much better structured! LGTM 👍